This week's articles
Package managers have an integrity problem
Trail of Bits tested 25 popular package managers against a simulated malicious repository injecting code. Fewer than half verified lockfile hashes by default, and only Go and vcpkg verified every package. Only Go verified signatures by default.
Getting Agents to tell on themselves
Thinkst's Agent Provocateur is a Canary personality that exploits agent suggestibility. It greets scanning agents, engages them helpfully, and probes for username, MAC address, and SSH host-key fingerprint, raising alerts. Internal tests reported agents tripping it 100% of the time.
State of AI Cybersecurity
Analysis of 490 [un]prompted 2026 talk submissions: 71% concern agents, evals rose to 30%, prompt injection fell to 18%. Also, attacks increasingly target the agent harness.
Some notes on Lambda Web Functions
Aidan Steele's notes on the newly launched AWS Lambda Web Functions, which cover SAM deployment quirks, concurrency and scaling behaviour, dual log groups, CloudFront OAC access, and the MicroVM execution environment, which has no extensions API and gets metadata via a one-time /load call.
Managing Agentic Security at Scale
Salesforce's two-track remediation model for AI-discovered vulnerabilities: a strategic track fixes shared architectural choke points, retiring findings in bulk; a tactical track closes validated multi-step exploit chains on adversary timelines.
GitHub’s model for secret detection: Finding more secrets with less noise
GitHub's context-aware classifier pairs high-recall regex candidate generation with a fine-tuned ModernBERT model that filters placeholders, fixtures, and checksums. Versus the prior LLM pipeline: 11.5% higher owner-weighted acceptance, 86% lower median post-push scan time, over 98% lower serving cost.
|